Unsupervised,corpus-based method for extending a biomedical terminology

26489

Authors

nan

Abstract

� Objectives:� To� automatically� extend� downwards�an�existing�biomedical�termi- nology� using� a� corpus� and� both� lexical� and�terminological�knowledge.�Methods:� Adjectival� modifiers� are� removed� from� terms� extracted� from� the� corpus� (three� million� noun� phrases� extracted� from� MEDLINE),� and� demodified� terms� are� searched� for� in� the� terminology� (UMLS� Metathesaurus,�restricted�to�disorders�and� procedures).� A� phrase� from� MEDLINE� becomes�a�candidate�term�in�the�Metathe- saurus�if�the�following�two�requirements� are� met:� 1)� a� demodified� term� created� from�this�phrase�is�found�in�the�terminol- ogy�and�2)�the�modifiers�removed�to�cre- ate� the� demodified� term� also� modify� existing�terms�from�the�terminology,�for�a� given� semantic� category.� A� manual� re- view�of�a�sample�of�candidate�terms�was� performed.�Results:�Out�of�the�3�million� simple� phrases� randomly� extracted� from� MEDLINE,� 125,000� new� terms� were� identified� for� inclusion� in� the� UMLS.� 83%�of�the�1000�terms�reviewed�manually� were� associated� with� a� relevant� UMLS� concept.� Discussion:� The� limitations� of� this�approach�are�discussed,�as�well�as�ad- aptation�and�generalization�issues.� 1�

Keywords

nan

Full Content

Unsupervised,�corpus-based�method�for�extending� a�biomedical�terminology� Olivier�Bodenreider� Thomas�C.�Rindflesch� Anita�Burgun� Lister�Hill�National�Center� for�Biomedical�Communications� National�Library�of�Medicine� Bethesda,�Maryland,�20894�-�USA� {olivier|tcr}@nlm.nih.gov� LIM�-�University�of�Rennes� Avenue�du�Pr�Léon�Bernard� 35043�Rennes�Cedex�-�France� Anita.Burgun� @univ-rennes1.fr� � � Abstract� Objectives:� To� automatically� extend� downwards�an�existing�biomedical�termi- nology� using� a� corpus� and� both� lexical� and�terminological�knowledge.�Methods:� Adjectival� modifiers� are� removed� from� terms� extracted� from� the� corpus� (three� million� noun� phrases� extracted� from� MEDLINE),� and� demodified� terms� are� searched� for� in� the� terminology� (UMLS� Metathesaurus,�restricted�to�disorders�and� procedures).� A� phrase� from� MEDLINE� becomes�a�candidate�term�in�the�Metathe- saurus�if�the�following�two�requirements� are� met:� 1)� a� demodified� term� created� from�this�phrase�is�found�in�the�terminol- ogy�and�2)�the�modifiers�removed�to�cre- ate� the� demodified� term� also� modify� existing�terms�from�the�terminology,�for�a� given� semantic� category.� A� manual� re- view�of�a�sample�of�candidate�terms�was� performed.�Results:�Out�of�the�3�million� simple� phrases� randomly� extracted� from� MEDLINE,� 125,000� new� terms� were� identified� for� inclusion� in� the� UMLS.� 83%�of�the�1000�terms�reviewed�manually� were� associated� with� a� relevant� UMLS� concept.� Discussion:� The� limitations� of� this�approach�are�discussed,�as�well�as�ad- aptation�and�generalization�issues.� 1� Introduction� Although� providing� a� reasonable� coverage� of� the� clinical� subdomain,� Chute� et� al.� (1996)� showed� that� terminological�resources�such�as�the�Interna- tional� Classificaton� of� Diseases,� SNOMED�Inter- national� or� the� UMLS� Metathesaurus� do� not� capture� all� the� concepts� needed� for� representing� clinical�concepts�in�patient�records.�In�a�subsequent� study,�Chute�and�Elkin�(1997)�suggested�that�quali- fiers� (including� adjectival�modifiers)�be�available� as�a�separate�axis,�in�order�to�both�increase�the�ex- pressivity� of� a� terminology� and� reduce� its� com- plexity�by�limiting�the�number�of�pre-coordinated� terms.� In�this�study,�rather�than�reducing�the�complex- ity,� we� use� modification� phenomena� in� order� to� investigate� a� corpus-based� methodology� for� auto- matically�discovering�new�terms�for�inclusion�in�a� controlled� vocabulary.� In�other�words,�our�objec- tive�is�to�acquire�hyponyms�for�terms�in�an�original� vocabulary�that�appear�in�the�literature�but�are�not� present�in�the�original�vocabulary.� 2� Background� Terms�play�a�major�role�in�a�variety�of�natural�lan- guage� processing� (NLP)� applications,� including� machine�translation,�text�understanding,�automatic� indexing,�and�information�retrieval.�Taking�advan- tage�of�the�availability�of�large�corpora,�automatic� terminology� acquisition�methods�were�developed,� for�example,�by�Bourigault�and�Jacquemin�(1999).� Word�affinities�generally�play�a�central�role�in� these� methods.� Grefenstette� (1994)� defines� three� Association for Computational Linguistics. the Biomedical Domain, Philadelphia, July 2002, pp. 53-60. Proceedings of the Workshop on Natural Language Processing in orders�of�word�affinities.�“First�order�affinities�de- scribe�collocates�of�words,�second-order�affinities� show�similarly�used�words,�and�third-order�affini- ties� create� semantic� groupings� among� similar� words”.� In� term� extraction,� this� analysis� is� often� applied�to�modifiers�in�order�to�establish�groups�of� terms�modified�by�a�given�modifier�or�the�list�of�all� possible�modifiers�for�a�given�term.� Hersh�et�al.�(1996)�demonstrated�the�feasibility� of�applying�natural�language�processing�techniques� to�a�corpus�of�clinical�narratives�from�an�electronic� medical�record�(EMR)�system.�Although�the�terms� extracted� were� compared� to� existing� terms�in�the� UMLS,�the�goal�of�this�study�was�vocabulary�dis- covery,�but�not�the�automatic�integration�of�newly� discovered�terms�into�the�terminology.� The�automatic�extension�of�an�existing�resource� based�on�a�corpus�has�also�been�studied.�For�ex- ample,�Habert�et�al.�(1998)�propose�a�method�for� extending�an�existing�specialized�semantic�lexicon.� Although�related�to�these�studies,�our�objective� is�to�automatically�extend�downwards�an�existing� biomedical�terminology�using�a�corpus�and�a�com- bination� of� lexical,� syntactic,� and� terminological� knowledge.� � In� this� study,� the� textual� source,� or� corpus,� is� MEDLINE®�1,�the�U.S.�National�Library�of�Medi- cine’s� (NLM)� premier� bibliographic� database.� MEDLINE�contains�over�eleven�million�references� to�articles�from�more�than�4,600�worldwide�jour- nals� in� life�sciences�with�a�concentration�on�bio- medicine.� We�use�the�Unified�Medical�Language�System®� (UMLS®)� Metathesaurus®�2� as� the� terminology�to� be� extended.� The� Metathesaurus,� also� developed� by� NLM,�is�organized�by�concept�or�meaning.�A� concept�is�defined�as�a�cluster�of�terms�represent- ing�the�same�meaning�(synonyms,�lexical�variants,� acronyms,� translations).� For� example,� names� for� the�disease�multiple�sclerosis�include�multiple�scle- rosis,�MS,�‘multiple�sclerosis,�NOS’,�disseminated� sclerosis,�and�sclérose�en�plaques.�The�13th�edition� (2002)�of�the�UMLS�Metathesaurus�contains�over� 1.5�million�unique�English�terms�drawn�from�more� than� sixty� families� of� medical� vocabularies,� and� organized�in�some�775,000�concepts.� ����������������������������������������������������������� 1�www.nlm.nih.gov/pubs/factsheets/medline.html� 2�umlsks.nlm.nih.gov� In� order� to� address� the� large� size� of� the� Metathesaurus,� we� limited� our� study� to� a� signifi- cant�subdomain�of�clinical�medicine:�disorders�and� procedures�(currently�about�615,000�unique�terms,� corresponding�to�some�157,000�disorder�concepts� and�95,000�medical�procedure�concepts).� In� the� UMLS,� each� concept� is� categorized� by� semantic� types� (ST)� from� the� semantic� network.� McCray� et� al.� (2001)� designed� groupings� of� STs� that� provide� a� partition� the� Metathesaurus� and,� therefore,�can�be�used�to�extract�consistent�sets�of� concepts� corresponding� to� a� subdomain,� such� as� disorders�or�procedures.� Disorder�and�procedure�terms�were�restricted�to� terms�suitable�for�natural�language�processing,�ex- cluding,�for�example,�such�terms�as�abdominal�in- jury,� NOS.� The� notation� “NOS”,� meaning� “not� otherwise�specified”,�is�a�marker�for�underspecifi- cation� often� found� in� terminological� resources.� When�identified�in�the�Metathesaurus,�obsolete�and� truncated� terms� were� also� excluded.� 477,491� unique�terms�were�selected�for�further�processing.� 3� Methods� The� approach� we� propose� for� discovering� candi- dates� for� Metathesaurus� concepts� is� to� compare� phrases� extracted� from� MEDLINE� to� current� UMLS� phrases.� We� capitalize� on� differences� in� modification� structure� between� the� MEDLINE� phrase�and�the�UMLS�phrase�to�determine�candi- dates�for�inclusion�in�the�Metathesaurus.�The�cru- cial� difference� is� between� a� phrase� containing� adjectival� modification� and� a� similar� phrase� “de- modified”�by�removing�its�adjectives.� A�phrase�from�MEDLINE�becomes�a�candidate� term�in�the�Metathesaurus�if�the�following�two�re- quirements�are�met:�1)�a�demodified�term�created� from�this�phrase�is�found�in�the�terminology�and�2)� similarly�modified�terms�exist�in�the�terminology,� for� a� given� semantic� category.� For� example,� the� phrase�pancreatic�bronchogenic�cyst�is�a�candidate� term� for� a�disorder�in�the�Metathesaurus�because� bronchogenic� cyst� exists� …

Content truncated to 500 words

Labeled Text

� Objectives:� To� automatically� extend� downwards�an�existing�biomedical�terminology� using� a� corpus� and� both� lexical� and�terminological�knowledge.�Methods:� Adjectival� modifiers� are� removed� from� terms� extracted� from� the� corpus� (three� million� noun� phrases� extracted� from� MEDLINE),� and� demodified� terms� are� searched� for� in� the� terminology� (UMLS� Metathesaurus,�restricted�to�disorders�and� procedures).� A� phrase� from� MEDLINE� becomes�a�candidate�term�in�the�Metathesaurus�if�the�following�two�requirements� are� met:� 1)� a� demodified� term� created� from�this�phrase�is�found�in�the�terminology�and�2)�the�modifiers�removed�to�create� the� demodified� term� also� modify� existing�terms�from�the�terminology,�for�a� given� semantic� category.� A� manual� review�of�a�sample�of�candidate�terms�was� performed.�Results:�Out�of�the�3�million� simple� phrases� randomly� extracted� from� MEDLINE,� 125,000� new� terms� were� identified� for� inclusion� in� the� UMLS.� 83%�of�the�1000�terms�reviewed�manually� were� associated� with� a� relevant� UMLS� concept.� Discussion:� The� limitations� of� this�approach�are�discussed,�as�well�as�adaptation�and�generalization�issues.� 1�

Publication Details

Paper ID:
26489
Year:
2002
Publication:
TIPSTER
Publication Type:
conf
DOI:
nan